ADR 0277 — A Telegram budget serves replies first and pauses on sends
Amends ADR 0274. Serves I9, I13.
Context
A consuming project moved every send of a bot onto one budget and read two properties that hurt a live
bot beside a campaign. The queue is first in, first out, so a reply to a person waits behind recipients:
at 24 messages a second, about 0.8 s behind 20 texts and about 8 s behind albums of 10 (calculated, not
measured). And run paused the whole token on any 429, including one on editMessageText of a single
progress card, so a limit on editing one message stopped the campaign.
Decision
The budget has two lanes over one rate. A call is interactive unless it says bulk. A waiting
interactive call is granted before any waiting bulk call; order inside a lane stays first in, first out.
The grants share one schedule, so the sum never exceeds messagesPerSecond, and a grant already made is
not revoked: a reply can still wait out the interval of the album before it.
A 429 pauses the token only when the method spends the budget. An edit or a read that sees one returns it
unchanged and pauses nothing. pauseAfter stays public for a caller who decides otherwise.
Alternatives
- Fair sharing between lanes: rejected because the point is that a person never waits on a campaign. A campaign has no deadline a reply can break, and a steady stream of replies is the bot working.
- A lane per subsystem: rejected. Two names answer the question that was asked, and more would be a scheduler nobody asked for.
Consequences
Sustained interactive traffic can starve bulk; that is the priority working, and stats().queuedBulk shows
it. A caller that relied on an edit’s 429 to stop sends now calls pauseAfter. Proven in
packages/core/tests/telegram-outbound-budget.test.ts.